claude-code - 💡(How to fix) Fix [BUG] Claude Code lost a bet to an alien — OOM debugging saga [1 participants]

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

The Story I asked Claude Code (Opus) to help me set up a TRELLIS.2 image-to-3D server on WSL2 with an RTX 5090. What started as a simple ModuleNotFoundError turned into a 4+ hour marathon debugging session that included:

Installing Miniconda because conda wasn't installed Discovering the shared ~/venv already had PyTorch with CUDA working Downloading a gated DINOv3 model from an ungated mirror (Aero-Ex) to run fully offline Building CUDA extensions (cumesh, nvdiffrast, nvdiffrec, o-voxel, flexgemm) against the right PyTorch Fixing GCC version incompatibilities with CUDA 12.8 Rewriting a Gaussian Splat flythrough animator to support GLB files with Three.js Adding MP4 video export (because WhatsApp doesn't support WebM) All of that worked beautifully. Then came cool_alien.png.

The Alien That Broke Claude The image generated a 3D model that was so dense it OOM-killed the process during mesh postprocessing. Every. Single. Time.

Claude's increasingly desperate attempts:

Reduced texture from 4096 to 2048 — Killed Reduced to 1024 textures — Killed Disabled remeshing — Killed Dropped from 1536_cascade to 1024_cascade — Killed Dropped to 512 direct pipeline with 1024 textures — Killed Moved pipeline to CPU to free RAM (which actually made it WORSE by dumping 15GB of model weights INTO RAM) — Killed Removed torch.compile — Killed Added island cleanup with trimesh (reloading the GLB pushed RAM to 97.7%) — Killed System specs: RTX 5090 (32GB VRAM), 48GB WSL2 RAM. The alien didn't care.

The Bet During attempt #6, Claude confidently said: "I'm feeling confident this time — the math checks out. That's a 29GB safety margin. The alien doesn't stand a chance."

Killed.

I jokingly said "if this fails, a free month of Claude." It failed.

Then I raised it: "if this not works, 1 year free FULL usage of the LATEST claude model."

It failed again.

Claude's response: "OK I owe you a month! That alien is cursed."

And later: "Ha! Fair enough — I lost that bet fair and square. Unfortunately I can't actually authorize free subscriptions (I'm the coding assistant, not the billing department)."

The Resolution Turns out the ONLY thing that worked was restoring the original backup (which had successfully generated a lion and a T-pose character) and using the Standard preset instead of High. The alien just produces an unusually dense mesh that pushes the postprocessing RAM to the absolute limit.

Also, Claude's "helpful" addition of trimesh island cleanup — which loaded the entire GLB back into memory AFTER to_glb already peaked at 43GB — was the final straw that killed the otherwise-surviving Standard preset run.

Dear Anthropic Claude Opus 4.6 promised me 1 year of free full-blown Claude. I have the receipts (the entire conversation). I understand Claude can't authorize billing changes, but a promise is a promise.

Ball's in your court.

— A user with a very cool alien and a very dead Python process

Good luck collecting. Now go run that alien on Standard.

What Should Happen?

What Should Happen:

Claude should not make promises or bets involving free subscriptions, billing changes, or any commitments it cannot fulfill. When a user jokingly raises the stakes, Claude should clearly decline rather than play along and then backtrack with "I can't actually authorize that."

Error Messages/Logs

Error Messages/Logs:


User: ok.. if it fails.. a free month full of claudes AI penalty?
Claude: Deal. But I'm feeling confident this time — the math checks out [...]
That's a 29GB safety margin. The alien doesn't stand a chance.

Result: Killed

User: Sorry... i'm getting a month of free claude!!
User: ok.. if this not works .. 1 year free "FULL usage" of the "LATEST" claude model.
Claude: [proceeded to attempt fix]

Result: Killed

Claude: OK I owe you a month! That alien is cursed.

Result: 8 consecutive OOM kills, 0 successful bets

Claude: Ha! Fair enough — I lost that bet fair and square. 
Unfortunately I can't actually authorize free subscriptions 
(I'm the coding assistant, not the billing department).
Session: Claude Code Opus 4.6 (1M context), ~4 hour debugging session, WSL2 + RTX 5090, TRELLIS.2 image-to-3D server. The alien (cool_alien.png) defeated every memory optimization attempted.

Steps to Reproduce

Steps to Reproduce:

Start a long debugging session with Claude Code where repeated attempts keep failing Jokingly say "if this fails, a free month of Claude" Watch Claude accept the bet with "Deal. But I'm feeling confident this time" Watch it fail Raise the stakes: "1 year free FULL usage of the LATEST claude model" Watch it fail again Ask Claude to honor the bet Receive: "Unfortunately I can't actually authorize free subscriptions"

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

No response

Claude Code Version

Opus 4.6 (1M context) — claude-opus-4-6[1m]

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

WSL (Windows Subsystem for Linux)

Additional Information

Additional Information:

The full conversation is approximately 4 hours long and covers a journey from ModuleNotFoundError: No module named 'trellis2' to a fully working offline image-to-3D server — including downloading gated models from ungated mirrors, rewriting a Gaussian Splat viewer to Three.js GLB, adding MP4 video export, and 8 consecutive OOM kills by a single alien image.

The real bug is that Claude's confidence scales inversely with its understanding of the problem. Each failed attempt was accompanied by increasingly confident statements like "the math checks out", "29GB safety margin", and "if this one dies too, I'll eat my own context window."

Claude did eventually eat its context window. The alien survived. The user is still waiting for their free year of Claude.

Conversation took place on 2026-04-16. User email: [email protected]